Reproduce “Reliable Robotic Task Execution in the Face of Anomalies” on OpenResearch

About this paper

Bharath Santhanam, Alex Mitrevski, Santosh Thoduka, Sebastian Houben, Teena Hassan

Researchers from NEURA Robotics, Chalmers University of Technology, Fraunhofer IAIS, and Hochschule Bonn-Rhein-Sieg developed a policy-agnostic framework for anomaly detection and recovery in learned robot policies. This framework uses visual cues to detect deviations from nominal execution and employs a multi-level recovery strategy, increasing task success rates from 0% to 75% for a dedicated policy on a door handle task and from 48% to 60% for a pre-trained Octo model on an object placing task in anomalous conditions.

Key insights

• Visual anomaly detection can be effectively achieved by learning features from nominal execution data and identifying states deviating from this learned distribution, without needing explicit failure examples.
• A structured, sequential recovery process, particularly one that can reset the robot to a new initial state based on a learned distribution of successful starting configurations, is crucial for escaping persistent anomalous situations.
• Decoupling the anomaly detection and recovery components from the core robot policy's architecture enables broad applicability across diverse learned policies, including large foundation models, without requiring their modification or retraining.

Summary by alphaXiv.